forked from EngineX-Cambricon/enginex-mlu370-vllm
50 lines
1.7 KiB
Python
50 lines
1.7 KiB
Python
"""Utilities for selecting and loading models."""
|
|
import contextlib
|
|
from typing import Tuple, Type
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from vllm.config import ModelConfig
|
|
from vllm.model_executor.models import ModelRegistry
|
|
|
|
|
|
@contextlib.contextmanager
|
|
def set_default_torch_dtype(dtype: torch.dtype):
|
|
"""Sets the default torch dtype to the given dtype."""
|
|
old_dtype = torch.get_default_dtype()
|
|
torch.set_default_dtype(dtype)
|
|
yield
|
|
torch.set_default_dtype(old_dtype)
|
|
|
|
|
|
def get_model_architecture(
|
|
model_config: ModelConfig) -> Tuple[Type[nn.Module], str]:
|
|
architectures = getattr(model_config.hf_config, "architectures", None) or []
|
|
print(f"[DEBUG-ARCH] get_model_architecture: "
|
|
f"type(hf_config)={type(model_config.hf_config).__name__}, "
|
|
f"architectures={getattr(model_config.hf_config, 'architectures', 'MISSING')}, "
|
|
f"id(hf_config)={id(model_config.hf_config)}")
|
|
# Special handling for quantized Mixtral.
|
|
# FIXME(woosuk): This is a temporary hack.
|
|
mixtral_supported = [
|
|
"fp8", "compressed-tensors", "gptq_marlin", "awq_marlin"
|
|
]
|
|
|
|
if (model_config.quantization is not None
|
|
and model_config.quantization not in mixtral_supported
|
|
and "MixtralForCausalLM" in architectures):
|
|
architectures = ["QuantMixtralForCausalLM"]
|
|
|
|
return ModelRegistry.resolve_model_cls(
|
|
architectures,
|
|
model_path=model_config.model,
|
|
revision=model_config.revision,
|
|
trust_remote_code=model_config.trust_remote_code,
|
|
hf_config=model_config.hf_config,
|
|
)
|
|
|
|
|
|
def get_architecture_class_name(model_config: ModelConfig) -> str:
|
|
return get_model_architecture(model_config)[1]
|